- matching to external resources
- string match on predicate is common way of linking to other resources
- references Zina Pozen’s work
- For 65% of things, pred matches work well
- For others, sense info in the pred is sometimes useful (look_v_like), sometimes not (story_n_of)
- does this generalize to: verb preds -> yes, noun preds -> no ?
- There are other subtypes of things that require idiosyncratic methods of matching
- There are many mismatches, e.g. sg vs pl distinctions (stair and stairs), which the ERG treats as the same, but WN does not
Discussion
FCB: anybody else trying to link resources
GS: thai language bi-lingual dictionaries, or rather, thai lemmas with eng definitions
DM: I’m linking the ERG to WN. I have done it before, but not to the extent or quality of analysis you’re proposing
RD: Not external resources, but I’m trying to link back to the surface form, as have Ned and Angelina….many interesting problems, though a small subset of your problems
FCB: **Our** problems
AC: How many items in WN have example sentences?
FCB: WN has example fragments (not sentences), for maybe a third. If you include semcor, etc., you get maybe 60%. Example sentences for a synset may have multiple lemmas.
TB: Do you annotate at the lexeme level or at the token level (parser output). E.g. with the chinese or japanese grammars, deverbal nouns etc., in a lexicon you might need an underspecified assignment it’s either a noun or a verb, or something else. What do you do with these?
FCB: We’re working at the token level and in context, but I expect the mapping to be more than a single predicate, so if we see a verbal noun and a light verb (e.g. suru in Jpn), we can use that.
LT: Are the senses for preds in the ERG ambiguous?
FCB: Yes, almost always
DM: Like “he drives a tank” or “he fills his tank”, tank are the same pred. Are you looking just at MRS or at the parse as well.
FCB: Just MRS
DM: So verb subcategorizations are in the MRS
FCB: In terms of verbal arguments, yes. So it’s possible that “tank of water” is a relational noun in the ERG and other tanks are not.
FCB: So looking at the mismatches, the interesting ones are things like “fasten” and “unfasten”, where the ERG treats it as compositional “un” + “fasten”
DM: Going back to the tank example, if you have verbnet, and the verb “drives” or “fills”, will that help disambiguate?
FCB: Disambiguation is another interesting question, but not what we’re doing now.
LT: Why do we have plural preds in the ERG?
DF: We have “he went into the woods”, where “woods” has nothing to do with the singular “wood”
LT: But if you put the lemma names from WN into the ERG, can you get something from that?
EMB: I think LT is saying taking info from WN, we can get useful information by putting it into the ERG, whereas the info is lost if we go the other way around
FCB: … What about morals? is that like moral? Isn’t the plural different than the singular?
DF: well that’s like “principles”, yes?
RD: Or “values” …
DF: If there is really a difference, then maybe the distinction belongs in the grammar.
oe: In […] paper we actually argue for the opposite.
FCB: The “morals” of this story are: it’s a difficult decision and we should be careful.
AC: This is a problem that lexicographers deal with all the time, and they don’t just decide this across the board. They sit down and talk about the cases as we do. If they chose to put the plural form in the dictionary, there’s a reason for it and we should take that reasoning seriously.
FCB: The question of “take advantage of” or even “kick the bucket” where WN is treating it as a single synset synonymous with “die”, but the ERG is compositional about it, this is where the idea of Schroedinger MRS (or Quantum MRS because we don’t want the cats to die) would be useful. What do people think… for “Kim kicked the bucket”, how many preds do we want?
oe: At [..] we turned on the idiom matching machinery, and I think that’s a good first step. We should make a decision about what we think is the right result when there’s an idiom.
…
FCB: A more compositional idiom, say “take advantage of”, you can do something for your benefit, or exploit, perhaps to someone’s harm.. Or “let the cat out of the bag”, where “cat” is the thing revealed, and “bag” is the secret or veil of obscurity. You can modify bits of it (TB: let the cat out of the velvet bag)
…
FCB: I had a student who found it very difficult to decide if idioms were decomposable
EMB: In Bellingham earlier this year, I was convinced of qMRS for compound names, but i don’t see that use case here
FCB: For “i lost my temper”, I’d like to have it also as “get angry”
EMB: why? Why do you need both?
FCB: Because I can lose my “hot temper”, or maybe for “lose my mind”, you can “lose your tiny mind” and so on, but it still means “go mad”, but the inner bits are modified.
EMB: As long as you’re changing the MRS, you can do coordination, e.g. “go crazy and mind is small”
FCB: But what about the normal form, do we have “go crazy and mind”? Or do we need idiom rules for the modified and the unmodified cases?
EMB: I still don’t feel the use cases you’re describing are the same as we discussed earlier
FCB: For names, yes, the use cases are different, but perhaps I’m not properly motivating the idiom case..
WP: If you have idioms with qMRS, why not other, non-idiomatic cases, like “love” and “adore”? This seems like slippery slope
FCB: Yes, we don’t want to link all synonyms, but when the paraphrases have different shape in the MRS.
…
FCB: These thing are facts of English, and they should be encoded. E.g. to lose one’s mind is not the same as to lose one’s wallet.
AC: We can do this with different senses of the verb, but I suppose I don’t have strong opinion at the moment about what to do here in general…
GS: We can say “kick the proverbial bucket”
AC: That’s a metalinguistic statement, and if we want to parse metalinguistic content then there’s a lot more to do
GS: “kick the freaking bucket”
…
FCB: Dan brought up sports metaphors, like “hit a home run”, which means to achieve success, or to literally hit a baseball outside the playing field, so the absolute minimum we want is to say there’s a hitting and homerun, but is this different than “take advantage of”? I don’t know how far I want to go with these, there’s always a large gray area of unclear cases. But in MT we don’t always want literal or idiomatic readings, it changes by context.
AC: If you’re doing a standard MT with alignment, you get the phrases that are translated, but if you’re working with the MRS, you get the ERG’s parse, which can make things harder
DF: But in phrases there can be distance that SMT would not capture, like “the tabs kept on sandy were closer than the ones kept on kim”, and in the MRS they are where they should be, but in the sentence the ngrams are quite far apart. I don’t think it’s true that the MRS is always doing damage to the signal.
AC: Maybe what we need are monolingual transfer rules that convert things to less idiomatic forms
RD: Aren’t we already doing this in MT with paraphrases?
AC: Yes, but we should probably be doing this in the greater delph-in ecosystem. I want it as part of the infrastructure, an obvious way to do it if we’re going to do it.
…
*applause*
Last update: 2014-07-15 by MichaelGoodman [edit]